%run "/code/source/notebooks/notebook_settings.py"
import logging
import numpy as np
import helpsk as hlp
from helpsk.utility import read_pickle, Timer
from helpsk.sklearn_eval import MLExperimentResults

import source.config.config as config
from source.service.model_registry import ModelRegistry

logging.info("Running experiment notebook for last run.")

2023-11-10 07:25:59 - INFO     | Running experiment notebook for last run.

Get Latest Experiment Run from MLFlow

registry = ModelRegistry(tracking_uri=config.experiment_server_url())
experiment = registry.get_experiment_by_name(exp_name=config.experiment_name())
logging.info(f"Experiment id: {experiment.last_run.exp_id}")
logging.info(f"Experiment name: {experiment.last_run.exp_name}")
logging.info(f"Run id: {experiment.last_run.run_id}")
logging.info(f"Metric(s): {experiment.last_run.metrics}")

2023-11-10 07:25:59 - INFO     | Experiment id: 1
2023-11-10 07:25:59 - INFO     | Experiment name: credit
2023-11-10 07:25:59 - INFO     | Run id: 505363f85fc741ada20bfb109bdee2d7
2023-11-10 07:25:59 - INFO     | Metric(s): {'roc_auc': 0.762641790401041}

Last Run vs Production

What is the metric/performance from the model associated with the last run?

logging.info(f"last run metrics: {experiment.last_run.metrics}")

2023-11-10 07:25:59 - INFO     | last run metrics: {'roc_auc': 0.762641790401041}

What is the metric/performance of the model in production?

production_run = registry.get_production_run(model_name=config.model_name())
logging.info(f"production run metrics: {production_run.metrics}")

2023-11-10 07:25:59 - INFO     | production run metrics: {'roc_auc': 0.7700760706900486}

Last Run

# underlying mlflow object
experiment.last_run.mlflow_entity

<Run: data=<RunData: metrics={'roc_auc': 0.762641790401041}, params={'model__criterion': 'entropy',
 'model__max_depth': '70',
 'model__max_features': '0.1142268477118407',
 'model__max_samples': '0.5483119512487002',
 'model__min_samples_leaf': '8',
 'model__min_samples_split': '12',
 'model__n_estimators': '553',
 'prep__non_numeric__encoder__transformer': "OneHotEncoder(handle_unknown='ignore')",
 'prep__numeric__imputer__transformer': "SimpleImputer(strategy='median')",
 'prep__numeric__pca__transformer': "PCA(n_components='mle')",
 'prep__numeric__scaler__transformer': 'None'}, tags={'mlflow.log-model.history': '[{"run_id": "505363f85fc741ada20bfb109bdee2d7", '
                             '"artifact_path": "model", "utc_time_created": '
                             '"2023-11-10 07:25:55.221442", "flavors": '
                             '{"python_function": {"model_path": "model.pkl", '
                             '"predict_fn": "predict", "loader_module": '
                             '"mlflow.sklearn", "python_version": "3.11.6", '
                             '"env": {"conda": "conda.yaml", "virtualenv": '
                             '"python_env.yaml"}}, "sklearn": '
                             '{"pickled_model": "model.pkl", '
                             '"sklearn_version": "1.3.2", '
                             '"serialization_format": "cloudpickle", "code": '
                             'null}}, "model_uuid": '
                             '"2821e66a22c54ba8b9630eef51fe20c6", '
                             '"mlflow_version": "2.8.0", "model_size_bytes": '
                             '2584736}]',
 'mlflow.note.content': '2023_11_10_07_25_30',
 'mlflow.runName': '2023_11_10_07_25_30',
 'mlflow.source.git.commit': 'e4a4afc62a952e1fc94d01102f96b08cd2840540',
 'mlflow.source.name': 'source/entrypoints/cli.py',
 'mlflow.source.type': 'LOCAL',
 'mlflow.user': 'root',
 'type': 'BayesSearchCV'}>, info=<RunInfo: artifact_uri='/code/mlflow-artifact-root/1/505363f85fc741ada20bfb109bdee2d7/artifacts', end_time=1699601156815, experiment_id='1', lifecycle_stage='active', run_id='505363f85fc741ada20bfb109bdee2d7', run_name='2023_11_10_07_25_30', run_uuid='505363f85fc741ada20bfb109bdee2d7', start_time=1699601130089, status='FINISHED', user_id='root'>, inputs=<RunInputs: dataset_inputs=[]>>

---

Load Training & Test Data Info

with Timer("Loading training/test datasets"):
    X_train = experiment.last_run.download_artifact(artifact_name='x_train.pkl', read_from=read_pickle)
    X_test = experiment.last_run.download_artifact(artifact_name='x_test.pkl', read_from=read_pickle)
    y_train = experiment.last_run.download_artifact(artifact_name='y_train.pkl', read_from=read_pickle)
    y_test = experiment.last_run.download_artifact(artifact_name='y_test.pkl', read_from=read_pickle)

Timer Started: Loading training/test datasets
Timer Finished (0.01 seconds)

logging.info(f"training X shape: {X_train.shape}")
logging.info(f"training y length: {len(y_train)}")

logging.info(f"test X shape: {X_test.shape}")
logging.info(f"test y length: {len(y_test)}")

2023-11-10 07:25:59 - INFO     | training X shape: (800, 20)
2023-11-10 07:25:59 - INFO     | training y length: 800
2023-11-10 07:25:59 - INFO     | test X shape: (200, 20)
2023-11-10 07:25:59 - INFO     | test y length: 200

np.unique(y_train, return_counts=True)

(array([0, 1]), array([559, 241]))

train_y_proportion = np.unique(y_train, return_counts=True)[1] \
    / np.sum(np.unique(y_train, return_counts=True)[1])
logging.info(f"balance of y in training: {train_y_proportion}")

2023-11-10 07:25:59 - INFO     | balance of y in training: [0.69875 0.30125]

test_y_proportion = np.unique(y_test, return_counts=True)[1] \
    / np.sum(np.unique(y_test, return_counts=True)[1])
logging.info(f"balance of y in test: {test_y_proportion}")

2023-11-10 07:25:59 - INFO     | balance of y in test: [0.705 0.295]

Cross Validation Results

Best Scores/Params

results = experiment.last_run.download_artifact(
    artifact_name='experiment.yaml',
    read_from=MLExperimentResults.from_yaml_file
)
logging.info(f"Best Score: {results.best_score}")
logging.info(f"Best Params: {results.best_params}")

2023-11-10 07:25:59 - INFO     | Best Score: 0.762641790401041
2023-11-10 07:25:59 - INFO     | Best Params: {'model': 'RandomForestClassifier()', 'max_features': 0.1142268477118407, 'max_depth': 70, 'n_estimators': 553, 'min_samples_split': 12, 'min_samples_leaf': 8, 'max_samples': 0.5483119512487002, 'criterion': 'entropy', 'imputer': "SimpleImputer(strategy='median')", 'scaler': 'None', 'pca': "PCA('mle')", 'encoder': 'OneHotEncoder()'}

# Best model from each model-type.
df = results.to_formatted_dataframe(return_style=False, include_rank=True)
df["model_rank"] = df.groupby("model")["roc_auc Mean"].rank(method="first", ascending=False)
df.query('model_rank == 1')

	rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	model	C	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	learning_rate	min_child_weight	subsample	colsample_bytree	colsample_bylevel	reg_alpha	reg_lambda	num_leaves	imputer	scaler	pca	encoder	model_rank
11	1	0.76	0.71	0.81	RandomForestClassifier()	NaN	0.11	70.00	553.00	12.00	8.00	0.55	entropy	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()	1.00
19	2	0.76	0.72	0.80	XGBClassifier()	NaN	NaN	1.00	896.00	NaN	NaN	NaN	NaN	0.03	8.00	0.80	0.91	0.83	0.00	1.41	NaN	SimpleImputer(strategy='median')	None	None	OneHotEncoder()	1.00
0	3	0.76	0.71	0.81	LogisticRegression()	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	StandardScaler()	None	OneHotEncoder()	1.00
24	4	0.76	0.73	0.79	LGBMClassifier()	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0.51	0.68	NaN	5.68	42.57	50.00	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()	1.00
9	6	0.75	0.70	0.81	ExtraTreesClassifier()	NaN	0.03	84.00	1088.00	24.00	36.00	0.98	gini	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	None	None	OneHotEncoder()	1.00

results.to_formatted_dataframe(return_style=True,
                               include_rank=True,
                               num_rows=500)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	model	C	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	learning_rate	min_child_weight	subsample	colsample_bytree	colsample_bylevel	reg_alpha	reg_lambda	num_leaves	imputer	scaler	pca	encoder
1	0.763	0.711	0.814	RandomForestClassifier()	<NA>	0.114	70.000	553.000	12.000	8.000	0.548	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
2	0.760	0.719	0.801	XGBClassifier()	<NA>	<NA>	1.000	896.000	<NA>	<NA>	<NA>	<NA>	0.029	8.000	0.799	0.906	0.825	0.003	1.411	<NA>	SimpleImputer(strategy='median')	None	None	OneHotEncoder()
3	0.759	0.713	0.805	LogisticRegression()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
4	0.757	0.726	0.788	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.511	0.683	<NA>	5.684	42.574	50.000	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
5	0.756	0.726	0.787	RandomForestClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
6	0.753	0.696	0.811	ExtraTreesClassifier()	<NA>	0.030	84.000	1,088.000	24.000	36.000	0.981	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
7	0.752	0.684	0.819	LogisticRegression()	0.001	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	MinMaxScaler()	None	OneHotEncoder()
8	0.750	0.711	0.789	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.797	0.700	<NA>	6.654	9.475	381.000	SimpleImputer(strategy='median')	None	None	CustomOrdinalEncoder()
9	0.744	0.708	0.780	RandomForestClassifier()	<NA>	0.681	38.000	1,461.000	23.000	10.000	0.553	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	CustomOrdinalEncoder()
10	0.743	0.690	0.795	ExtraTreesClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
11	0.739	0.670	0.807	LogisticRegression()	23.327	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	StandardScaler()	None	OneHotEncoder()
12	0.738	0.705	0.772	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
13	0.737	0.696	0.779	RandomForestClassifier()	<NA>	0.710	15.000	1,493.000	33.000	27.000	0.914	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	None	PCA('mle')	OneHotEncoder()
14	0.731	0.712	0.750	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.598	0.661	<NA>	12.533	35.084	348.000	SimpleImputer(strategy='most_frequent')	None	None	CustomOrdinalEncoder()
15	0.729	0.664	0.794	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.845	0.453	<NA>	16.166	40.978	351.000	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
16	0.727	0.670	0.785	XGBClassifier()	<NA>	<NA>	5.000	1,218.000	<NA>	<NA>	<NA>	<NA>	0.115	2.000	0.545	0.648	0.852	0.123	1.165	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	CustomOrdinalEncoder()
17	0.726	0.689	0.762	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	StandardScaler()	None	CustomOrdinalEncoder()
18	0.725	0.699	0.752	RandomForestClassifier()	<NA>	0.740	14.000	1,645.000	5.000	43.000	0.741	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	None	PCA('mle')	CustomOrdinalEncoder()
19	0.723	0.694	0.753	XGBClassifier()	<NA>	<NA>	15.000	1,159.000	<NA>	<NA>	<NA>	<NA>	0.032	29.000	0.834	0.520	0.503	0.003	1.839	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	CustomOrdinalEncoder()
20	0.723	0.666	0.779	ExtraTreesClassifier()	<NA>	0.857	30.000	879.000	17.000	28.000	0.563	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	CustomOrdinalEncoder()
21	0.722	0.655	0.790	ExtraTreesClassifier()	<NA>	0.672	81.000	1,136.000	34.000	34.000	0.971	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	CustomOrdinalEncoder()
22	0.722	0.684	0.760	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	StandardScaler()	PCA('mle')	CustomOrdinalEncoder()
23	0.722	0.658	0.786	ExtraTreesClassifier()	<NA>	0.781	50.000	590.000	35.000	47.000	0.846	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
24	0.720	0.706	0.733	XGBClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
25	0.714	0.666	0.763	XGBClassifier()	<NA>	<NA>	3.000	682.000	<NA>	<NA>	<NA>	<NA>	0.152	2.000	0.698	0.940	0.817	0.009	2.086	<NA>	SimpleImputer()	None	PCA('mle')	CustomOrdinalEncoder()

results.to_formatted_dataframe(query='model == "RandomForestClassifier()"', include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	imputer	pca	encoder
1	0.763	0.711	0.814	0.114	70.000	553.000	12.000	8.000	0.548	entropy	SimpleImputer(strategy='median')	PCA('mle')	OneHotEncoder()
2	0.756	0.726	0.787	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	OneHotEncoder()
3	0.744	0.708	0.780	0.681	38.000	1,461.000	23.000	10.000	0.553	gini	SimpleImputer(strategy='median')	None	CustomOrdinalEncoder()
4	0.737	0.696	0.779	0.710	15.000	1,493.000	33.000	27.000	0.914	gini	SimpleImputer(strategy='most_frequent')	PCA('mle')	OneHotEncoder()
5	0.725	0.699	0.752	0.740	14.000	1,645.000	5.000	43.000	0.741	entropy	SimpleImputer(strategy='most_frequent')	PCA('mle')	CustomOrdinalEncoder()

results.to_formatted_dataframe(query='model == "LogisticRegression()"', include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	C	imputer	scaler	pca	encoder
1	0.759	0.713	0.805	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
2	0.752	0.684	0.819	0.001	SimpleImputer(strategy='median')	MinMaxScaler()	None	OneHotEncoder()
3	0.739	0.670	0.807	23.327	SimpleImputer(strategy='median')	StandardScaler()	None	OneHotEncoder()
4	0.726	0.689	0.762	0.000	SimpleImputer(strategy='median')	StandardScaler()	None	CustomOrdinalEncoder()
5	0.722	0.684	0.760	0.000	SimpleImputer(strategy='median')	StandardScaler()	PCA('mle')	CustomOrdinalEncoder()

BayesSearchCV Performance Over Time

results.plot_performance_across_trials(facet_by='model').show()

results.plot_performance_across_trials(query='model == "RandomForestClassifier()"').show()

---

Variable Performance Over Time

results.plot_parameter_values_across_trials(query='model == "RandomForestClassifier()"').show()

---

Scatter Matrix

# results.plot_scatter_matrix(query='model == "RandomForestClassifier()"',
#                             height=1000, width=1000).show()

---

Variable Performance - Numeric

results.plot_performance_numeric_params(query='model == "RandomForestClassifier()"',
                                        height=800)

results.plot_parallel_coordinates(query='model == "RandomForestClassifier()"').show()

---

Variable Performance - Non-Numeric

results.plot_performance_non_numeric_params(query='model == "RandomForestClassifier()"').show()

---

results.plot_score_vs_parameter(
    query='model == "RandomForestClassifier()"',
    parameter='max_features',
    size='max_depth',
    color='encoder',
)

---

# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='max_depth'
# )

# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='imputer'
# )

---

Last Run - Test Set Performance

last_model = experiment.last_run.download_artifact(
    artifact_name='model/model.pkl',
    read_from=read_pickle
)
print(type(last_model.model))

<class 'sklearn.pipeline.Pipeline'>

last_model

SklearnModelWrapper(model=Pipeline(steps=[('prep',
                                           ColumnTransformer(transformers=[('numeric',
                                                                            Pipeline(steps=[('imputer',
                                                                                             TransformerChooser(transformer=SimpleImputer(strategy='median'))),
                                                                                            ('scaler',
                                                                                             TransformerChooser()),
                                                                                            ('pca',
                                                                                             TransformerChooser(transformer=PCA(n_components='mle')))]),
                                                                            ['duration',
                                                                             'credit_amount',
                                                                             'installment_commitment',
                                                                             'residence_si...
                                                                             'savings_status',
                                                                             'employment',
                                                                             'personal_status',
                                                                             'other_parties',
                                                                             'property_magnitude',
                                                                             'other_payment_plans',
                                                                             'housing',
                                                                             'job',
                                                                             'own_telephone',
                                                                             'foreign_worker'])])),
                                          ('model',
                                           RandomForestClassifier(criterion='entropy',
                                                                  max_depth=70,
                                                                  max_features=0.1142268477118407,
                                                                  max_samples=0.5483119512487002,
                                                                  min_samples_leaf=8,
                                                                  min_samples_split=12,
                                                                  n_estimators=553,
                                                                  random_state=42))]))

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

test_predictions = last_model.predict(X_test)
test_predictions[0:10]

array([0.3733136 , 0.4118129 , 0.48404516, 0.35906406, 0.17344127,
       0.33008237, 0.17795812, 0.40549351, 0.21201897, 0.24426042])

evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37
)

evaluator.plot_actual_vs_predict_histogram()

evaluator.plot_confusion_matrix()

evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)

	Score	Dummy (prior)	Dummy (constant)	Explanation
AUC	0.794	0.500	0.500	Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate	0.593	0.000	1.000	59.3% of positive instances were correctly identified.; i.e. 35 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate	0.837	1.000	0.000	83.7% of negative instances were correctly identified.; i.e. 118 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate	0.163	0.000	1.000	16.3% of negative instances were incorrectly identified as positive; i.e. 23 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate	0.407	1.000	0.000	40.7% of positive instances were incorrectly identified as negative; i.e. 24 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value	0.603	0.000	0.295	When the model claims an instance is positive, it is correct 60.3% of the time; i.e. out of the 58 times the model predicted "Positive Class", it was correct 35 times; a.k.a precision
Negative Predictive Value	0.831	0.705	0.000	When the model claims an instance is negative, it is correct 83.1% of the time; i.e. out of the 142 times the model predicted "Negative Class", it was correct 118 times
F1 Score	0.598	0.000	0.456	The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC	0.642	0.295	0.295	Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy	0.765	0.705	0.295	76.5% of instances were correctly identified
Error Rate	0.235	0.295	0.705	23.5% of instances were incorrectly identified
% Positive	0.295	0.295	0.295	29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations	200	200	200	There are 200 total observations; i.e. sample size

evaluator.plot_roc_auc_curve().show()

evaluator.plot_precision_recall_auc_curve().show()

evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()

evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()

evaluator.calculate_lift_gain(return_style=True)

/usr/local/lib/python3.11/site-packages/helpsk/sklearn_eval.py:2480: FutureWarning:

The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.

	Gain	Lift
Percentile
5	0.17	3.39
10	0.25	2.54
15	0.32	2.15
20	0.42	2.12
25	0.53	2.10
30	0.61	2.03
35	0.68	1.94
40	0.75	1.86
45	0.76	1.69
50	0.80	1.59
55	0.80	1.45
60	0.88	1.47
65	0.88	1.36
70	0.92	1.31
75	0.93	1.24
80	0.97	1.21
85	0.98	1.16
90	1.00	1.11
95	1.00	1.05
100	1.00	1.00

---

Production Model - Test Set Performance

production_model = production_run.download_artifact(
    artifact_name='model/model.pkl',
    read_from=read_pickle
)
print(type(production_model.model))

<class 'sklearn.pipeline.Pipeline'>

production_model

SklearnModelWrapper(model=Pipeline(steps=[('prep',
                                           ColumnTransformer(transformers=[('numeric',
                                                                            Pipeline(steps=[('imputer',
                                                                                             TransformerChooser(transformer=SimpleImputer(strategy='median'))),
                                                                                            ('scaler',
                                                                                             TransformerChooser()),
                                                                                            ('pca',
                                                                                             TransformerChooser(transformer=PCA(n_components='mle')))]),
                                                                            ['duration',
                                                                             'credit_amount',
                                                                             'installment_commitment',
                                                                             'residence_si...
                                                                             'personal_status',
                                                                             'other_parties',
                                                                             'property_magnitude',
                                                                             'other_payment_plans',
                                                                             'housing',
                                                                             'job',
                                                                             'own_telephone',
                                                                             'foreign_worker'])])),
                                          ('model',
                                           ExtraTreesClassifier(bootstrap=True,
                                                                criterion='entropy',
                                                                max_depth=99,
                                                                max_features=0.031837350792579364,
                                                                max_samples=0.9248344222191298,
                                                                min_samples_leaf=4,
                                                                min_samples_split=16,
                                                                n_estimators=1235,
                                                                random_state=42))]))

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

test_predictions = production_model.predict(X_test)
test_predictions[0:10]

array([0.32794716, 0.32893009, 0.37430263, 0.3052219 , 0.23870015,
       0.31131838, 0.24759746, 0.33864031, 0.23092148, 0.24922654])

evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37
)

evaluator.plot_actual_vs_predict_histogram()

evaluator.plot_confusion_matrix()

evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)

	Score	Dummy (prior)	Dummy (constant)	Explanation
AUC	0.781	0.500	0.500	Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate	0.169	0.000	1.000	16.9% of positive instances were correctly identified.; i.e. 10 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate	0.957	1.000	0.000	95.7% of negative instances were correctly identified.; i.e. 135 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate	0.043	0.000	1.000	4.3% of negative instances were incorrectly identified as positive; i.e. 6 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate	0.831	1.000	0.000	83.1% of positive instances were incorrectly identified as negative; i.e. 49 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value	0.625	0.000	0.295	When the model claims an instance is positive, it is correct 62.5% of the time; i.e. out of the 16 times the model predicted "Positive Class", it was correct 10 times; a.k.a precision
Negative Predictive Value	0.734	0.705	0.000	When the model claims an instance is negative, it is correct 73.4% of the time; i.e. out of the 184 times the model predicted "Negative Class", it was correct 135 times
F1 Score	0.267	0.000	0.456	The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC	0.621	0.295	0.295	Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy	0.725	0.705	0.295	72.5% of instances were correctly identified
Error Rate	0.275	0.295	0.705	27.5% of instances were incorrectly identified
% Positive	0.295	0.295	0.295	29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations	200	200	200	There are 200 total observations; i.e. sample size

evaluator.plot_roc_auc_curve().show()

evaluator.plot_precision_recall_auc_curve().show()

evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()

evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()

evaluator.calculate_lift_gain(return_style=True)

/usr/local/lib/python3.11/site-packages/helpsk/sklearn_eval.py:2480: FutureWarning:

The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.

	Gain	Lift
Percentile
5	0.17	3.39
10	0.20	2.03
15	0.34	2.26
20	0.44	2.20
25	0.49	1.97
30	0.56	1.86
35	0.63	1.79
40	0.69	1.74
45	0.76	1.69
50	0.80	1.59
55	0.83	1.51
60	0.85	1.41
65	0.86	1.33
70	0.93	1.33
75	0.95	1.27
80	0.97	1.21
85	0.98	1.16
90	1.00	1.11
95	1.00	1.05
100	1.00	1.00

---